Imagined Value Gradients: Model-Based Policy Optimization with Transferable Latent Dynamics Models